kernel/net/ipv4/tcp.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  *
  20  * Fixes:
  21  *              Alan Cox        :       Numerous verify_area() calls
  22  *              Alan Cox        :       Set the ACK bit on a reset
  23  *              Alan Cox        :       Stopped it crashing if it closed while
  24  *                                      sk->inuse=1 and was trying to connect
  25  *                                      (tcp_err()).
  26  *              Alan Cox        :       All icmp error handling was broken
  27  *                                      pointers passed where wrong and the
  28  *                                      socket was looked up backwards. Nobody
  29  *                                      tested any icmp error code obviously.
  30  *              Alan Cox        :       tcp_err() now handled properly. It
  31  *                                      wakes people on errors. poll
  32  *                                      behaves and the icmp error race
  33  *                                      has gone by moving it into sock.c
  34  *              Alan Cox        :       tcp_send_reset() fixed to work for
  35  *                                      everything not just packets for
  36  *                                      unknown sockets.
  37  *              Alan Cox        :       tcp option processing.
  38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
  39  *                                      syn rule wrong]
  40  *              Herp Rosmanith  :       More reset fixes
  41  *              Alan Cox        :       No longer acks invalid rst frames.
  42  *                                      Acking any kind of RST is right out.
  43  *              Alan Cox        :       Sets an ignore me flag on an rst
  44  *                                      receive otherwise odd bits of prattle
  45  *                                      escape still
  46  *              Alan Cox        :       Fixed another acking RST frame bug.
  47  *                                      Should stop LAN workplace lockups.
  48  *              Alan Cox        :       Some tidyups using the new skb list
  49  *                                      facilities
  50  *              Alan Cox        :       sk->keepopen now seems to work
  51  *              Alan Cox        :       Pulls options out correctly on accepts
  52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
  53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
  54  *                                      bit to skb ops.
  55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
  56  *                                      nasty.
  57  *              Alan Cox        :       Added some better commenting, as the
  58  *                                      tcp is hard to follow
  59  *              Alan Cox        :       Removed incorrect check for 20 * psh
  60  *      Michael O'Reilly        :       ack < copied bug fix.
  61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
  62  *              Alan Cox        :       FIN with no memory -> CRASH
  63  *              Alan Cox        :       Added socket option proto entries.
  64  *                                      Also added awareness of them to accept.
  65  *              Alan Cox        :       Added TCP options (SOL_TCP)
  66  *              Alan Cox        :       Switched wakeup calls to callbacks,
  67  *                                      so the kernel can layer network
  68  *                                      sockets.
  69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
  70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
  71  *              Alan Cox        :       RST frames sent on unsynchronised
  72  *                                      state ack error.
  73  *              Alan Cox        :       Put in missing check for SYN bit.
  74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
  75  *                                      window non shrink trick.
  76  *              Alan Cox        :       Added a couple of small NET2E timer
  77  *                                      fixes
  78  *              Charles Hedrick :       TCP fixes
  79  *              Toomas Tamm     :       TCP window fixes
  80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
  81  *              Charles Hedrick :       Rewrote most of it to actually work
  82  *              Linus           :       Rewrote tcp_read() and URG handling
  83  *                                      completely
  84  *              Gerhard Koerting:       Fixed some missing timer handling
  85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
  86  *              Gerhard Koerting:       PC/TCP workarounds
  87  *              Adam Caldwell   :       Assorted timer/timing errors
  88  *              Matthew Dillon  :       Fixed another RST bug
  89  *              Alan Cox        :       Move to kernel side addressing changes.
  90  *              Alan Cox        :       Beginning work on TCP fastpathing
  91  *                                      (not yet usable)
  92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
  93  *              Alan Cox        :       TCP fast path debugging
  94  *              Alan Cox        :       Window clamping
  95  *              Michael Riepe   :       Bug in tcp_check()
  96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
  97  *              Matt Dillon     :       Yet more small nasties remove from the
  98  *                                      TCP code (Be very nice to this man if
  99  *                                      tcp finally works 100%) 8)
 100  *              Alan Cox        :       BSD accept semantics.
 101  *              Alan Cox        :       Reset on closedown bug.
 102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
 103  *              Michael Pall    :       Handle poll() after URG properly in
 104  *                                      all cases.
 105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
 106  *                                      (multi URG PUSH broke rlogin).
 107  *              Michael Pall    :       Fix the multi URG PUSH problem in
 108  *                                      tcp_readable(), poll() after URG
 109  *                                      works now.
 110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
 111  *                                      BSD api.
 112  *              Alan Cox        :       Changed the semantics of sk->socket to
 113  *                                      fix a race and a signal problem with
 114  *                                      accept() and async I/O.
 115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
 116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
 117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
 118  *                                      clients/servers which listen in on
 119  *                                      fixed ports.
 120  *              Alan Cox        :       Cleaned the above up and shrank it to
 121  *                                      a sensible code size.
 122  *              Alan Cox        :       Self connect lockup fix.
 123  *              Alan Cox        :       No connect to multicast.
 124  *              Ross Biro       :       Close unaccepted children on master
 125  *                                      socket close.
 126  *              Alan Cox        :       Reset tracing code.
 127  *              Alan Cox        :       Spurious resets on shutdown.
 128  *              Alan Cox        :       Giant 15 minute/60 second timer error
 129  *              Alan Cox        :       Small whoops in polling before an
 130  *                                      accept.
 131  *              Alan Cox        :       Kept the state trace facility since
 132  *                                      it's handy for debugging.
 133  *              Alan Cox        :       More reset handler fixes.
 134  *              Alan Cox        :       Started rewriting the code based on
 135  *                                      the RFC's for other useful protocol
 136  *                                      references see: Comer, KA9Q NOS, and
 137  *                                      for a reference on the difference
 138  *                                      between specifications and how BSD
 139  *                                      works see the 4.4lite source.
 140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
 141  *                                      close.
 142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
 143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
 144  *              Alan Cox        :       Reimplemented timers as per the RFC
 145  *                                      and using multiple timers for sanity.
 146  *              Alan Cox        :       Small bug fixes, and a lot of new
 147  *                                      comments.
 148  *              Alan Cox        :       Fixed dual reader crash by locking
 149  *                                      the buffers (much like datagram.c)
 150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
 151  *                                      now gets fed up of retrying without
 152  *                                      (even a no space) answer.
 153  *              Alan Cox        :       Extracted closing code better
 154  *              Alan Cox        :       Fixed the closing state machine to
 155  *                                      resemble the RFC.
 156  *              Alan Cox        :       More 'per spec' fixes.
 157  *              Jorge Cwik      :       Even faster checksumming.
 158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
 159  *                                      only frames. At least one pc tcp stack
 160  *                                      generates them.
 161  *              Alan Cox        :       Cache last socket.
 162  *              Alan Cox        :       Per route irtt.
 163  *              Matt Day        :       poll()->select() match BSD precisely on error
 164  *              Alan Cox        :       New buffers
 165  *              Marc Tamsky     :       Various sk->prot->retransmits and
 166  *                                      sk->retransmits misupdating fixed.
 167  *                                      Fixed tcp_write_timeout: stuck close,
 168  *                                      and TCP syn retries gets used now.
 169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
 170  *                                      ack if state is TCP_CLOSED.
 171  *              Alan Cox        :       Look up device on a retransmit - routes may
 172  *                                      change. Doesn't yet cope with MSS shrink right
 173  *                                      but it's a start!
 174  *              Marc Tamsky     :       Closing in closing fixes.
 175  *              Mike Shaver     :       RFC1122 verifications.
 176  *              Alan Cox        :       rcv_saddr errors.
 177  *              Alan Cox        :       Block double connect().
 178  *              Alan Cox        :       Small hooks for enSKIP.
 179  *              Alexey Kuznetsov:       Path MTU discovery.
 180  *              Alan Cox        :       Support soft errors.
 181  *              Alan Cox        :       Fix MTU discovery pathological case
 182  *                                      when the remote claims no mtu!
 183  *              Marc Tamsky     :       TCP_CLOSE fix.
 184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
 185  *                                      window but wrong (fixes NT lpd problems)
 186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
 187  *              Joerg Reuter    :       No modification of locked buffers in
 188  *                                      tcp_do_retransmit()
 189  *              Eric Schenk     :       Changed receiver side silly window
 190  *                                      avoidance algorithm to BSD style
 191  *                                      algorithm. This doubles throughput
 192  *                                      against machines running Solaris,
 193  *                                      and seems to result in general
 194  *                                      improvement.
 195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
 196  *      Willy Konynenberg       :       Transparent proxying support.
 197  *      Mike McLagan            :       Routing by source
 198  *              Keith Owens     :       Do proper merging with partial SKB's in
 199  *                                      tcp_do_sendmsg to avoid burstiness.
 200  *              Eric Schenk     :       Fix fast close down bug with
 201  *                                      shutdown() followed by close().
 202  *              Andi Kleen      :       Make poll agree with SIGIO
 203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
 204  *                                      lingertime == 0 (RFC 793 ABORT Call)
 205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
 206  *                                      csum_and_copy_from_user() if possible.
 207  *
 208  *              This program is free software; you can redistribute it and/or
 209  *              modify it under the terms of the GNU General Public License
 210  *              as published by the Free Software Foundation; either version
 211  *              2 of the License, or(at your option) any later version.
 212  *
 213  * Description of States:
 214  *
 215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
 216  *
 217  *      TCP_SYN_RECV            received a connection request, sent ack,
 218  *                              waiting for final ack in three-way handshake.
 219  *
 220  *      TCP_ESTABLISHED         connection established
 221  *
 222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
 223  *                              transmission of remaining buffered data
 224  *
 225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
 226  *                              to shutdown
 227  *
 228  *      TCP_CLOSING             both sides have shutdown but we still have
 229  *                              data we have to finish sending
 230  *
 231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
 232  *                              closed, can only be entered from FIN_WAIT2
 233  *                              or CLOSING.  Required because the other end
 234  *                              may not have gotten our last ACK causing it
 235  *                              to retransmit the data packet (which we ignore)
 236  *
 237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
 238  *                              us to finish writing our data and to shutdown
 239  *                              (we have to close() to move on to LAST_ACK)
 240  *
 241  *      TCP_LAST_ACK            out side has shutdown after remote has
 242  *                              shutdown.  There may still be data in our
 243  *                              buffer that we have to finish sending
 244  *
 245  *      TCP_CLOSE               socket is finished
 246  */
 247
 248 #define pr_fmt(fmt) "TCP: " fmt
 249
 250 #include <linux/kernel.h>
 251 #include <linux/module.h>
 252 #include <linux/types.h>
 253 #include <linux/fcntl.h>
 254 #include <linux/poll.h>
 255 #include <linux/inet_diag.h>
 256 #include <linux/init.h>
 257 #include <linux/fs.h>
 258 #include <linux/skbuff.h>
 259 #include <linux/scatterlist.h>
 260 #include <linux/splice.h>
 261 #include <linux/net.h>
 262 #include <linux/socket.h>
 263 #include <linux/random.h>
 264 #include <linux/bootmem.h>
 265 #include <linux/highmem.h>
 266 #include <linux/swap.h>
 267 #include <linux/cache.h>
 268 #include <linux/err.h>
 269 #include <linux/crypto.h>
 270 #include <linux/time.h>
 271 #include <linux/slab.h>
 272
 273 #include <net/icmp.h>
 274 #include <net/inet_common.h>
 275 #include <net/tcp.h>
 276 #include <net/xfrm.h>
 277 #include <net/ip.h>
 278 #include <net/sock.h>
 279
 280 #include <asm/uaccess.h>
 281 #include <asm/ioctls.h>
 282 #include <net/busy_poll.h>
 283
 284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 285
 286 int sysctl_tcp_min_tso_segs __read_mostly = 2;
 287
 288 int sysctl_tcp_autocorking __read_mostly = 1;
 289
 290 struct percpu_counter tcp_orphan_count;
 291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 292
 293 long sysctl_tcp_mem[3] __read_mostly;
 294 int sysctl_tcp_wmem[3] __read_mostly;
 295 int sysctl_tcp_rmem[3] __read_mostly;
 296
 297 EXPORT_SYMBOL(sysctl_tcp_mem);
 298 EXPORT_SYMBOL(sysctl_tcp_rmem);
 299 EXPORT_SYMBOL(sysctl_tcp_wmem);
 300
 301 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
 302 EXPORT_SYMBOL(tcp_memory_allocated);
 303
 304 /*
 305  * Current number of TCP sockets.
 306  */
 307 struct percpu_counter tcp_sockets_allocated;
 308 EXPORT_SYMBOL(tcp_sockets_allocated);
 309
 310 /*
 311  * TCP splice context
 312  */
 313 struct tcp_splice_state {
 314         struct pipe_inode_info *pipe;
 315         size_t len;
 316         unsigned int flags;
 317 };
 318
 319 /*
 320  * Pressure flag: try to collapse.
 321  * Technical note: it is used by multiple contexts non atomically.
 322  * All the __sk_mem_schedule() is of this nature: accounting
 323  * is strict, actions are advisory and have some latency.
 324  */
 325 int tcp_memory_pressure __read_mostly;
 326 EXPORT_SYMBOL(tcp_memory_pressure);
 327
 328 void tcp_enter_memory_pressure(struct sock *sk)
 329 {
 330         if (!tcp_memory_pressure) {
 331                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 332                 tcp_memory_pressure = 1;
 333         }
 334 }
 335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
 336
 337 /* Convert seconds to retransmits based on initial and max timeout */
 338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 339 {
 340         u8 res = 0;
 341
 342         if (seconds > 0) {
 343                 int period = timeout;
 344
 345                 res = 1;
 346                 while (seconds > period && res < 255) {
 347                         res++;
 348                         timeout <<= 1;
 349                         if (timeout > rto_max)
 350                                 timeout = rto_max;
 351                         period += timeout;
 352                 }
 353         }
 354         return res;
 355 }
 356
 357 /* Convert retransmits to seconds based on initial and max timeout */
 358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 359 {
 360         int period = 0;
 361
 362         if (retrans > 0) {
 363                 period = timeout;
 364                 while (--retrans) {
 365                         timeout <<= 1;
 366                         if (timeout > rto_max)
 367                                 timeout = rto_max;
 368                         period += timeout;
 369                 }
 370         }
 371         return period;
 372 }
 373
 374 /* Address-family independent initialization for a tcp_sock.
 375  *
 376  * NOTE: A lot of things set to zero explicitly by call to
 377  *       sk_alloc() so need not be done here.
 378  */
 379 void tcp_init_sock(struct sock *sk)
 380 {
 381         struct inet_connection_sock *icsk = inet_csk(sk);
 382         struct tcp_sock *tp = tcp_sk(sk);
 383
 384         __skb_queue_head_init(&tp->out_of_order_queue);
 385         tcp_init_xmit_timers(sk);
 386         tcp_prequeue_init(tp);
 387         INIT_LIST_HEAD(&tp->tsq_node);
 388
 389         icsk->icsk_rto = TCP_TIMEOUT_INIT;
 390         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 391
 392         /* So many TCP implementations out there (incorrectly) count the
 393          * initial SYN frame in their delayed-ACK and congestion control
 394          * algorithms that we must have the following bandaid to talk
 395          * efficiently to them.  -DaveM
 396          */
 397         tp->snd_cwnd = TCP_INIT_CWND;
 398
 399         /* See draft-stevens-tcpca-spec-01 for discussion of the
 400          * initialization of these values.
 401          */
 402         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 403         tp->snd_cwnd_clamp = ~0;
 404         tp->mss_cache = TCP_MSS_DEFAULT;
 405         u64_stats_init(&tp->syncp);
 406
 407         tp->reordering = sysctl_tcp_reordering;
 408         tcp_enable_early_retrans(tp);
 409         tcp_assign_congestion_control(sk);
 410
 411         tp->tsoffset = 0;
 412
 413         sk->sk_state = TCP_CLOSE;
 414
 415         sk->sk_write_space = sk_stream_write_space;
 416         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 417
 418         icsk->icsk_sync_mss = tcp_sync_mss;
 419
 420         sk->sk_sndbuf = sysctl_tcp_wmem[1];
 421         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
 422
 423         local_bh_disable();
 424         sock_update_memcg(sk);
 425         sk_sockets_allocated_inc(sk);
 426         local_bh_enable();
 427 }
 428 EXPORT_SYMBOL(tcp_init_sock);
 429
 430 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
 431 {
 432         if (sk->sk_tsflags) {
 433                 struct skb_shared_info *shinfo = skb_shinfo(skb);
 434
 435                 sock_tx_timestamp(sk, &shinfo->tx_flags);
 436                 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
 437                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 438         }
 439 }
 440
 441 /*
 442  *      Wait for a TCP event.
 443  *
 444  *      Note that we don't need to lock the socket, as the upper poll layers
 445  *      take care of normal races (between the test and the event) and we don't
 446  *      go look at any of the socket buffers directly.
 447  */
 448 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 449 {
 450         unsigned int mask;
 451         struct sock *sk = sock->sk;
 452         const struct tcp_sock *tp = tcp_sk(sk);
 453
 454         sock_rps_record_flow(sk);
 455
 456         sock_poll_wait(file, sk_sleep(sk), wait);
 457         if (sk->sk_state == TCP_LISTEN)
 458                 return inet_csk_listen_poll(sk);
 459
 460         /* Socket is not locked. We are protected from async events
 461          * by poll logic and correct handling of state changes
 462          * made by other threads is impossible in any case.
 463          */
 464
 465         mask = 0;
 466
 467         /*
 468          * POLLHUP is certainly not done right. But poll() doesn't
 469          * have a notion of HUP in just one direction, and for a
 470          * socket the read side is more interesting.
 471          *
 472          * Some poll() documentation says that POLLHUP is incompatible
 473          * with the POLLOUT/POLLWR flags, so somebody should check this
 474          * all. But careful, it tends to be safer to return too many
 475          * bits than too few, and you can easily break real applications
 476          * if you don't tell them that something has hung up!
 477          *
 478          * Check-me.
 479          *
 480          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
 481          * our fs/select.c). It means that after we received EOF,
 482          * poll always returns immediately, making impossible poll() on write()
 483          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
 484          * if and only if shutdown has been made in both directions.
 485          * Actually, it is interesting to look how Solaris and DUX
 486          * solve this dilemma. I would prefer, if POLLHUP were maskable,
 487          * then we could set it on SND_SHUTDOWN. BTW examples given
 488          * in Stevens' books assume exactly this behaviour, it explains
 489          * why POLLHUP is incompatible with POLLOUT.    --ANK
 490          *
 491          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 492          * blocking on fresh not-connected or disconnected socket. --ANK
 493          */
 494         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
 495                 mask |= POLLHUP;
 496         if (sk->sk_shutdown & RCV_SHUTDOWN)
 497                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
 498
 499         /* Connected or passive Fast Open socket? */
 500         if (sk->sk_state != TCP_SYN_SENT &&
 501             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) {
 502                 int target = sock_rcvlowat(sk, 0, INT_MAX);
 503
 504                 if (tp->urg_seq == tp->copied_seq &&
 505                     !sock_flag(sk, SOCK_URGINLINE) &&
 506                     tp->urg_data)
 507                         target++;
 508
 509                 /* Potential race condition. If read of tp below will
 510                  * escape above sk->sk_state, we can be illegally awaken
 511                  * in SYN_* states. */
 512                 if (tp->rcv_nxt - tp->copied_seq >= target)
 513                         mask |= POLLIN | POLLRDNORM;
 514
 515                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 516                         if (sk_stream_is_writeable(sk)) {
 517                                 mask |= POLLOUT | POLLWRNORM;
 518                         } else {  /* send SIGIO later */
 519                                 set_bit(SOCK_ASYNC_NOSPACE,
 520                                         &sk->sk_socket->flags);
 521                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 522
 523                                 /* Race breaker. If space is freed after
 524                                  * wspace test but before the flags are set,
 525                                  * IO signal will be lost. Memory barrier
 526                                  * pairs with the input side.
 527                                  */
 528                                 smp_mb__after_atomic();
 529                                 if (sk_stream_is_writeable(sk))
 530                                         mask |= POLLOUT | POLLWRNORM;
 531                         }
 532                 } else
 533                         mask |= POLLOUT | POLLWRNORM;
 534
 535                 if (tp->urg_data & TCP_URG_VALID)
 536                         mask |= POLLPRI;
 537         }
 538         /* This barrier is coupled with smp_wmb() in tcp_reset() */
 539         smp_rmb();
 540         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 541                 mask |= POLLERR;
 542
 543         return mask;
 544 }
 545 EXPORT_SYMBOL(tcp_poll);
 546
 547 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 548 {
 549         struct tcp_sock *tp = tcp_sk(sk);
 550         int answ;
 551         bool slow;
 552
 553         switch (cmd) {
 554         case SIOCINQ:
 555                 if (sk->sk_state == TCP_LISTEN)
 556                         return -EINVAL;
 557
 558                 slow = lock_sock_fast(sk);
 559                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 560                         answ = 0;
 561                 else if (sock_flag(sk, SOCK_URGINLINE) ||
 562                          !tp->urg_data ||
 563                          before(tp->urg_seq, tp->copied_seq) ||
 564                          !before(tp->urg_seq, tp->rcv_nxt)) {
 565
 566                         answ = tp->rcv_nxt - tp->copied_seq;
 567
 568                         /* Subtract 1, if FIN was received */
 569                         if (answ && sock_flag(sk, SOCK_DONE))
 570                                 answ--;
 571                 } else
 572                         answ = tp->urg_seq - tp->copied_seq;
 573                 unlock_sock_fast(sk, slow);
 574                 break;
 575         case SIOCATMARK:
 576                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
 577                 break;
 578         case SIOCOUTQ:
 579                 if (sk->sk_state == TCP_LISTEN)
 580                         return -EINVAL;
 581
 582                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 583                         answ = 0;
 584                 else
 585                         answ = tp->write_seq - tp->snd_una;
 586                 break;
 587         case SIOCOUTQNSD:
 588                 if (sk->sk_state == TCP_LISTEN)
 589                         return -EINVAL;
 590
 591                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 592                         answ = 0;
 593                 else
 594                         answ = tp->write_seq - tp->snd_nxt;
 595                 break;
 596         default:
 597                 return -ENOIOCTLCMD;
 598         }
 599
 600         return put_user(answ, (int __user *)arg);
 601 }
 602 EXPORT_SYMBOL(tcp_ioctl);
 603
 604 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 605 {
 606         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 607         tp->pushed_seq = tp->write_seq;
 608 }
 609
 610 static inline bool forced_push(const struct tcp_sock *tp)
 611 {
 612         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 613 }
 614
 615 static void skb_entail(struct sock *sk, struct sk_buff *skb)
 616 {
 617         struct tcp_sock *tp = tcp_sk(sk);
 618         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 619
 620         skb->csum    = 0;
 621         tcb->seq     = tcb->end_seq = tp->write_seq;
 622         tcb->tcp_flags = TCPHDR_ACK;
 623         tcb->sacked  = 0;
 624         __skb_header_release(skb);
 625         tcp_add_write_queue_tail(sk, skb);
 626         sk->sk_wmem_queued += skb->truesize;
 627         sk_mem_charge(sk, skb->truesize);
 628         if (tp->nonagle & TCP_NAGLE_PUSH)
 629                 tp->nonagle &= ~TCP_NAGLE_PUSH;
 630 }
 631
 632 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 633 {
 634         if (flags & MSG_OOB)
 635                 tp->snd_up = tp->write_seq;
 636 }
 637
 638 /* If a not yet filled skb is pushed, do not send it if
 639  * we have data packets in Qdisc or NIC queues :
 640  * Because TX completion will happen shortly, it gives a chance
 641  * to coalesce future sendmsg() payload into this skb, without
 642  * need for a timer, and with no latency trade off.
 643  * As packets containing data payload have a bigger truesize
 644  * than pure acks (dataless) packets, the last checks prevent
 645  * autocorking if we only have an ACK in Qdisc/NIC queues,
 646  * or if TX completion was delayed after we processed ACK packet.
 647  */
 648 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 649                                 int size_goal)
 650 {
 651         return skb->len < size_goal &&
 652                sysctl_tcp_autocorking &&
 653                skb != tcp_write_queue_head(sk) &&
 654                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
 655 }
 656
 657 static void tcp_push(struct sock *sk, int flags, int mss_now,
 658                      int nonagle, int size_goal)
 659 {
 660         struct tcp_sock *tp = tcp_sk(sk);
 661         struct sk_buff *skb;
 662
 663         if (!tcp_send_head(sk))
 664                 return;
 665
 666         skb = tcp_write_queue_tail(sk);
 667         if (!(flags & MSG_MORE) || forced_push(tp))
 668                 tcp_mark_push(tp, skb);
 669
 670         tcp_mark_urg(tp, flags);
 671
 672         if (tcp_should_autocork(sk, skb, size_goal)) {
 673
 674                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
 675                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
 676                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 677                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
 678                 }
 679                 /* It is possible TX completion already happened
 680                  * before we set TSQ_THROTTLED.
 681                  */
 682                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
 683                         return;
 684         }
 685
 686         if (flags & MSG_MORE)
 687                 nonagle = TCP_NAGLE_CORK;
 688
 689         __tcp_push_pending_frames(sk, mss_now, nonagle);
 690 }
 691
 692 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 693                                 unsigned int offset, size_t len)
 694 {
 695         struct tcp_splice_state *tss = rd_desc->arg.data;
 696         int ret;
 697
 698         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
 699                               tss->flags);
 700         if (ret > 0)
 701                 rd_desc->count -= ret;
 702         return ret;
 703 }
 704
 705 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 706 {
 707         /* Store TCP splice context information in read_descriptor_t. */
 708         read_descriptor_t rd_desc = {
 709                 .arg.data = tss,
 710                 .count    = tss->len,
 711         };
 712
 713         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 714 }
 715
 716 /**
 717  *  tcp_splice_read - splice data from TCP socket to a pipe
 718  * @sock:       socket to splice from
 719  * @ppos:       position (not valid)
 720  * @pipe:       pipe to splice to
 721  * @len:        number of bytes to splice
 722  * @flags:      splice modifier flags
 723  *
 724  * Description:
 725  *    Will read pages from given socket and fill them into a pipe.
 726  *
 727  **/
 728 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 729                         struct pipe_inode_info *pipe, size_t len,
 730                         unsigned int flags)
 731 {
 732         struct sock *sk = sock->sk;
 733         struct tcp_splice_state tss = {
 734                 .pipe = pipe,
 735                 .len = len,
 736                 .flags = flags,
 737         };
 738         long timeo;
 739         ssize_t spliced;
 740         int ret;
 741
 742         sock_rps_record_flow(sk);
 743         /*
 744          * We can't seek on a socket input
 745          */
 746         if (unlikely(*ppos))
 747                 return -ESPIPE;
 748
 749         ret = spliced = 0;
 750
 751         lock_sock(sk);
 752
 753         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 754         while (tss.len) {
 755                 ret = __tcp_splice_read(sk, &tss);
 756                 if (ret < 0)
 757                         break;
 758                 else if (!ret) {
 759                         if (spliced)
 760                                 break;
 761                         if (sock_flag(sk, SOCK_DONE))
 762                                 break;
 763                         if (sk->sk_err) {
 764                                 ret = sock_error(sk);
 765                                 break;
 766                         }
 767                         if (sk->sk_shutdown & RCV_SHUTDOWN)
 768                                 break;
 769                         if (sk->sk_state == TCP_CLOSE) {
 770                                 /*
 771                                  * This occurs when user tries to read
 772                                  * from never connected socket.
 773                                  */
 774                                 if (!sock_flag(sk, SOCK_DONE))
 775                                         ret = -ENOTCONN;
 776                                 break;
 777                         }
 778                         if (!timeo) {
 779                                 ret = -EAGAIN;
 780                                 break;
 781                         }
 782                         sk_wait_data(sk, &timeo);
 783                         if (signal_pending(current)) {
 784                                 ret = sock_intr_errno(timeo);
 785                                 break;
 786                         }
 787                         continue;
 788                 }
 789                 tss.len -= ret;
 790                 spliced += ret;
 791
 792                 if (!timeo)
 793                         break;
 794                 release_sock(sk);
 795                 lock_sock(sk);
 796
 797                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 798                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
 799                     signal_pending(current))
 800                         break;
 801         }
 802
 803         release_sock(sk);
 804
 805         if (spliced)
 806                 return spliced;
 807
 808         return ret;
 809 }
 810 EXPORT_SYMBOL(tcp_splice_read);
 811
 812 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 813 {
 814         struct sk_buff *skb;
 815
 816         /* The TCP header must be at least 32-bit aligned.  */
 817         size = ALIGN(size, 4);
 818
 819         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 820         if (skb) {
 821                 if (sk_wmem_schedule(sk, skb->truesize)) {
 822                         skb_reserve(skb, sk->sk_prot->max_header);
 823                         /*
 824                          * Make sure that we have exactly size bytes
 825                          * available to the caller, no more, no less.
 826                          */
 827                         skb->reserved_tailroom = skb->end - skb->tail - size;
 828                         return skb;
 829                 }
 830                 __kfree_skb(skb);
 831         } else {
 832                 sk->sk_prot->enter_memory_pressure(sk);
 833                 sk_stream_moderate_sndbuf(sk);
 834         }
 835         return NULL;
 836 }
 837
 838 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 839                                        int large_allowed)
 840 {
 841         struct tcp_sock *tp = tcp_sk(sk);
 842         u32 new_size_goal, size_goal;
 843
 844         if (!large_allowed || !sk_can_gso(sk))
 845                 return mss_now;
 846
 847         /* Note : tcp_tso_autosize() will eventually split this later */
 848         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 849         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 850
 851         /* We try hard to avoid divides here */
 852         size_goal = tp->gso_segs * mss_now;
 853         if (unlikely(new_size_goal < size_goal ||
 854                      new_size_goal >= size_goal + mss_now)) {
 855                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 856                                      sk->sk_gso_max_segs);
 857                 size_goal = tp->gso_segs * mss_now;
 858         }
 859
 860         return max(size_goal, mss_now);
 861 }
 862
 863 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 864 {
 865         int mss_now;
 866
 867         mss_now = tcp_current_mss(sk);
 868         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 869
 870         return mss_now;
 871 }
 872
 873 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 874                                 size_t size, int flags)
 875 {
 876         struct tcp_sock *tp = tcp_sk(sk);
 877         int mss_now, size_goal;
 878         int err;
 879         ssize_t copied;
 880         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 881
 882         /* Wait for a connection to finish. One exception is TCP Fast Open
 883          * (passive side) where data is allowed to be sent before a connection
 884          * is fully established.
 885          */
 886         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 887             !tcp_passive_fastopen(sk)) {
 888                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
 889                         goto out_err;
 890         }
 891
 892         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 893
 894         mss_now = tcp_send_mss(sk, &size_goal, flags);
 895         copied = 0;
 896
 897         err = -EPIPE;
 898         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 899                 goto out_err;
 900
 901         while (size > 0) {
 902                 struct sk_buff *skb = tcp_write_queue_tail(sk);
 903                 int copy, i;
 904                 bool can_coalesce;
 905
 906                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
 907 new_segment:
 908                         if (!sk_stream_memory_free(sk))
 909                                 goto wait_for_sndbuf;
 910
 911                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
 912                         if (!skb)
 913                                 goto wait_for_memory;
 914
 915                         skb_entail(sk, skb);
 916                         copy = size_goal;
 917                 }
 918
 919                 if (copy > size)
 920                         copy = size;
 921
 922                 i = skb_shinfo(skb)->nr_frags;
 923                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
 924                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
 925                         tcp_mark_push(tp, skb);
 926                         goto new_segment;
 927                 }
 928                 if (!sk_wmem_schedule(sk, copy))
 929                         goto wait_for_memory;
 930
 931                 if (can_coalesce) {
 932                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 933                 } else {
 934                         get_page(page);
 935                         skb_fill_page_desc(skb, i, page, offset, copy);
 936                 }
 937                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 938
 939                 skb->len += copy;
 940                 skb->data_len += copy;
 941                 skb->truesize += copy;
 942                 sk->sk_wmem_queued += copy;
 943                 sk_mem_charge(sk, copy);
 944                 skb->ip_summed = CHECKSUM_PARTIAL;
 945                 tp->write_seq += copy;
 946                 TCP_SKB_CB(skb)->end_seq += copy;
 947                 tcp_skb_pcount_set(skb, 0);
 948
 949                 if (!copied)
 950                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 951
 952                 copied += copy;
 953                 offset += copy;
 954                 if (!(size -= copy)) {
 955                         tcp_tx_timestamp(sk, skb);
 956                         goto out;
 957                 }
 958
 959                 if (skb->len < size_goal || (flags & MSG_OOB))
 960                         continue;
 961
 962                 if (forced_push(tp)) {
 963                         tcp_mark_push(tp, skb);
 964                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
 965                 } else if (skb == tcp_send_head(sk))
 966                         tcp_push_one(sk, mss_now);
 967                 continue;
 968
 969 wait_for_sndbuf:
 970                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 971 wait_for_memory:
 972                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
 973                          TCP_NAGLE_PUSH, size_goal);
 974
 975                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
 976                         goto do_error;
 977
 978                 mss_now = tcp_send_mss(sk, &size_goal, flags);
 979         }
 980
 981 out:
 982         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
 983                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
 984         return copied;
 985
 986 do_error:
 987         if (copied)
 988                 goto out;
 989 out_err:
 990         return sk_stream_error(sk, flags, err);
 991 }
 992
 993 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
 994                  size_t size, int flags)
 995 {
 996         ssize_t res;
 997
 998         if (!(sk->sk_route_caps & NETIF_F_SG) ||
 999             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1000                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1001                                         flags);
1002
1003         lock_sock(sk);
1004         res = do_tcp_sendpages(sk, page, offset, size, flags);
1005         release_sock(sk);
1006         return res;
1007 }
1008 EXPORT_SYMBOL(tcp_sendpage);
1009
1010 static inline int select_size(const struct sock *sk, bool sg)
1011 {
1012         const struct tcp_sock *tp = tcp_sk(sk);
1013         int tmp = tp->mss_cache;
1014
1015         if (sg) {
1016                 if (sk_can_gso(sk)) {
1017                         /* Small frames wont use a full page:
1018                          * Payload will immediately follow tcp header.
1019                          */
1020                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1021                 } else {
1022                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1023
1024                         if (tmp >= pgbreak &&
1025                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1026                                 tmp = pgbreak;
1027                 }
1028         }
1029
1030         return tmp;
1031 }
1032
1033 void tcp_free_fastopen_req(struct tcp_sock *tp)
1034 {
1035         if (tp->fastopen_req) {
1036                 kfree(tp->fastopen_req);
1037                 tp->fastopen_req = NULL;
1038         }
1039 }
1040
1041 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1042                                 int *copied, size_t size)
1043 {
1044         struct tcp_sock *tp = tcp_sk(sk);
1045         int err, flags;
1046
1047         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1048                 return -EOPNOTSUPP;
1049         if (tp->fastopen_req)
1050                 return -EALREADY; /* Another Fast Open is in progress */
1051
1052         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1053                                    sk->sk_allocation);
1054         if (unlikely(!tp->fastopen_req))
1055                 return -ENOBUFS;
1056         tp->fastopen_req->data = msg;
1057         tp->fastopen_req->size = size;
1058
1059         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1060         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1061                                     msg->msg_namelen, flags);
1062         *copied = tp->fastopen_req->copied;
1063         tcp_free_fastopen_req(tp);
1064         return err;
1065 }
1066
1067 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1068 {
1069         struct tcp_sock *tp = tcp_sk(sk);
1070         struct sk_buff *skb;
1071         int flags, err, copied = 0;
1072         int mss_now = 0, size_goal, copied_syn = 0;
1073         bool sg;
1074         long timeo;
1075
1076         lock_sock(sk);
1077
1078         flags = msg->msg_flags;
1079         if (flags & MSG_FASTOPEN) {
1080                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1081                 if (err == -EINPROGRESS && copied_syn > 0)
1082                         goto out;
1083                 else if (err)
1084                         goto out_err;
1085         }
1086
1087         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1088
1089         /* Wait for a connection to finish. One exception is TCP Fast Open
1090          * (passive side) where data is allowed to be sent before a connection
1091          * is fully established.
1092          */
1093         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1094             !tcp_passive_fastopen(sk)) {
1095                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1096                         goto do_error;
1097         }
1098
1099         if (unlikely(tp->repair)) {
1100                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1101                         copied = tcp_send_rcvq(sk, msg, size);
1102                         goto out_nopush;
1103                 }
1104
1105                 err = -EINVAL;
1106                 if (tp->repair_queue == TCP_NO_QUEUE)
1107                         goto out_err;
1108
1109                 /* 'common' sending to sendq */
1110         }
1111
1112         /* This should be in poll */
1113         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1114
1115         mss_now = tcp_send_mss(sk, &size_goal, flags);
1116
1117         /* Ok commence sending. */
1118         copied = 0;
1119
1120         err = -EPIPE;
1121         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1122                 goto out_err;
1123
1124         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1125
1126         while (msg_data_left(msg)) {
1127                 int copy = 0;
1128                 int max = size_goal;
1129
1130                 skb = tcp_write_queue_tail(sk);
1131                 if (tcp_send_head(sk)) {
1132                         if (skb->ip_summed == CHECKSUM_NONE)
1133                                 max = mss_now;
1134                         copy = max - skb->len;
1135                 }
1136
1137                 if (copy <= 0) {
1138 new_segment:
1139                         /* Allocate new segment. If the interface is SG,
1140                          * allocate skb fitting to single page.
1141                          */
1142                         if (!sk_stream_memory_free(sk))
1143                                 goto wait_for_sndbuf;
1144
1145                         skb = sk_stream_alloc_skb(sk,
1146                                                   select_size(sk, sg),
1147                                                   sk->sk_allocation);
1148                         if (!skb)
1149                                 goto wait_for_memory;
1150
1151                         /*
1152                          * Check whether we can use HW checksum.
1153                          */
1154                         if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1155                                 skb->ip_summed = CHECKSUM_PARTIAL;
1156
1157                         skb_entail(sk, skb);
1158                         copy = size_goal;
1159                         max = size_goal;
1160
1161                         /* All packets are restored as if they have
1162                          * already been sent. skb_mstamp isn't set to
1163                          * avoid wrong rtt estimation.
1164                          */
1165                         if (tp->repair)
1166                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1167                 }
1168
1169                 /* Try to append data to the end of skb. */
1170                 if (copy > msg_data_left(msg))
1171                         copy = msg_data_left(msg);
1172
1173                 /* Where to copy to? */
1174                 if (skb_availroom(skb) > 0) {
1175                         /* We have some space in skb head. Superb! */
1176                         copy = min_t(int, copy, skb_availroom(skb));
1177                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1178                         if (err)
1179                                 goto do_fault;
1180                 } else {
1181                         bool merge = true;
1182                         int i = skb_shinfo(skb)->nr_frags;
1183                         struct page_frag *pfrag = sk_page_frag(sk);
1184
1185                         if (!sk_page_frag_refill(sk, pfrag))
1186                                 goto wait_for_memory;
1187
1188                         if (!skb_can_coalesce(skb, i, pfrag->page,
1189                                               pfrag->offset)) {
1190                                 if (i == MAX_SKB_FRAGS || !sg) {
1191                                         tcp_mark_push(tp, skb);
1192                                         goto new_segment;
1193                                 }
1194                                 merge = false;
1195                         }
1196
1197                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1198
1199                         if (!sk_wmem_schedule(sk, copy))
1200                                 goto wait_for_memory;
1201
1202                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1203                                                        pfrag->page,
1204                                                        pfrag->offset,
1205                                                        copy);
1206                         if (err)
1207                                 goto do_error;
1208
1209                         /* Update the skb. */
1210                         if (merge) {
1211                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1212                         } else {
1213                                 skb_fill_page_desc(skb, i, pfrag->page,
1214                                                    pfrag->offset, copy);
1215                                 get_page(pfrag->page);
1216                         }
1217                         pfrag->offset += copy;
1218                 }
1219
1220                 if (!copied)
1221                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1222
1223                 tp->write_seq += copy;
1224                 TCP_SKB_CB(skb)->end_seq += copy;
1225                 tcp_skb_pcount_set(skb, 0);
1226
1227                 copied += copy;
1228                 if (!msg_data_left(msg)) {
1229                         tcp_tx_timestamp(sk, skb);
1230                         goto out;
1231                 }
1232
1233                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1234                         continue;
1235
1236                 if (forced_push(tp)) {
1237                         tcp_mark_push(tp, skb);
1238                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1239                 } else if (skb == tcp_send_head(sk))
1240                         tcp_push_one(sk, mss_now);
1241                 continue;
1242
1243 wait_for_sndbuf:
1244                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1245 wait_for_memory:
1246                 if (copied)
1247                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1248                                  TCP_NAGLE_PUSH, size_goal);
1249
1250                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1251                         goto do_error;
1252
1253                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1254         }
1255
1256 out:
1257         if (copied)
1258                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1259 out_nopush:
1260         release_sock(sk);
1261         return copied + copied_syn;
1262
1263 do_fault:
1264         if (!skb->len) {
1265                 tcp_unlink_write_queue(skb, sk);
1266                 /* It is the one place in all of TCP, except connection
1267                  * reset, where we can be unlinking the send_head.
1268                  */
1269                 tcp_check_send_head(sk, skb);
1270                 sk_wmem_free_skb(sk, skb);
1271         }
1272
1273 do_error:
1274         if (copied + copied_syn)
1275                 goto out;
1276 out_err:
1277         err = sk_stream_error(sk, flags, err);
1278         release_sock(sk);
1279         return err;
1280 }
1281 EXPORT_SYMBOL(tcp_sendmsg);
1282
1283 /*
1284  *      Handle reading urgent data. BSD has very simple semantics for
1285  *      this, no blocking and very strange errors 8)
1286  */
1287
1288 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1289 {
1290         struct tcp_sock *tp = tcp_sk(sk);
1291
1292         /* No URG data to read. */
1293         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1294             tp->urg_data == TCP_URG_READ)
1295                 return -EINVAL; /* Yes this is right ! */
1296
1297         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1298                 return -ENOTCONN;
1299
1300         if (tp->urg_data & TCP_URG_VALID) {
1301                 int err = 0;
1302                 char c = tp->urg_data;
1303
1304                 if (!(flags & MSG_PEEK))
1305                         tp->urg_data = TCP_URG_READ;
1306
1307                 /* Read urgent data. */
1308                 msg->msg_flags |= MSG_OOB;
1309
1310                 if (len > 0) {
1311                         if (!(flags & MSG_TRUNC))
1312                                 err = memcpy_to_msg(msg, &c, 1);
1313                         len = 1;
1314                 } else
1315                         msg->msg_flags |= MSG_TRUNC;
1316
1317                 return err ? -EFAULT : len;
1318         }
1319
1320         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1321                 return 0;
1322
1323         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1324          * the available implementations agree in this case:
1325          * this call should never block, independent of the
1326          * blocking state of the socket.
1327          * Mike <pall@rz.uni-karlsruhe.de>
1328          */
1329         return -EAGAIN;
1330 }
1331
1332 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1333 {
1334         struct sk_buff *skb;
1335         int copied = 0, err = 0;
1336
1337         /* XXX -- need to support SO_PEEK_OFF */
1338
1339         skb_queue_walk(&sk->sk_write_queue, skb) {
1340                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1341                 if (err)
1342                         break;
1343
1344                 copied += skb->len;
1345         }
1346
1347         return err ?: copied;
1348 }
1349
1350 /* Clean up the receive buffer for full frames taken by the user,
1351  * then send an ACK if necessary.  COPIED is the number of bytes
1352  * tcp_recvmsg has given to the user so far, it speeds up the
1353  * calculation of whether or not we must ACK for the sake of
1354  * a window update.
1355  */
1356 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1357 {
1358         struct tcp_sock *tp = tcp_sk(sk);
1359         bool time_to_ack = false;
1360
1361         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1362
1363         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1364              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1365              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1366
1367         if (inet_csk_ack_scheduled(sk)) {
1368                 const struct inet_connection_sock *icsk = inet_csk(sk);
1369                    /* Delayed ACKs frequently hit locked sockets during bulk
1370                     * receive. */
1371                 if (icsk->icsk_ack.blocked ||
1372                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1373                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1374                     /*
1375                      * If this read emptied read buffer, we send ACK, if
1376                      * connection is not bidirectional, user drained
1377                      * receive buffer and there was a small segment
1378                      * in queue.
1379                      */
1380                     (copied > 0 &&
1381                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1382                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1383                        !icsk->icsk_ack.pingpong)) &&
1384                       !atomic_read(&sk->sk_rmem_alloc)))
1385                         time_to_ack = true;
1386         }
1387
1388         /* We send an ACK if we can now advertise a non-zero window
1389          * which has been raised "significantly".
1390          *
1391          * Even if window raised up to infinity, do not send window open ACK
1392          * in states, where we will not receive more. It is useless.
1393          */
1394         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1395                 __u32 rcv_window_now = tcp_receive_window(tp);
1396
1397                 /* Optimize, __tcp_select_window() is not cheap. */
1398                 if (2*rcv_window_now <= tp->window_clamp) {
1399                         __u32 new_window = __tcp_select_window(sk);
1400
1401                         /* Send ACK now, if this read freed lots of space
1402                          * in our buffer. Certainly, new_window is new window.
1403                          * We can advertise it now, if it is not less than current one.
1404                          * "Lots" means "at least twice" here.
1405                          */
1406                         if (new_window && new_window >= 2 * rcv_window_now)
1407                                 time_to_ack = true;
1408                 }
1409         }
1410         if (time_to_ack)
1411                 tcp_send_ack(sk);
1412 }
1413
1414 static void tcp_prequeue_process(struct sock *sk)
1415 {
1416         struct sk_buff *skb;
1417         struct tcp_sock *tp = tcp_sk(sk);
1418
1419         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1420
1421         /* RX process wants to run with disabled BHs, though it is not
1422          * necessary */
1423         local_bh_disable();
1424         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1425                 sk_backlog_rcv(sk, skb);
1426         local_bh_enable();
1427
1428         /* Clear memory counter. */
1429         tp->ucopy.memory = 0;
1430 }
1431
1432 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1433 {
1434         struct sk_buff *skb;
1435         u32 offset;
1436
1437         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1438                 offset = seq - TCP_SKB_CB(skb)->seq;
1439                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1440                         offset--;
1441                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1442                         *off = offset;
1443                         return skb;
1444                 }
1445                 /* This looks weird, but this can happen if TCP collapsing
1446                  * splitted a fat GRO packet, while we released socket lock
1447                  * in skb_splice_bits()
1448                  */
1449                 sk_eat_skb(sk, skb);
1450         }
1451         return NULL;
1452 }
1453
1454 /*
1455  * This routine provides an alternative to tcp_recvmsg() for routines
1456  * that would like to handle copying from skbuffs directly in 'sendfile'
1457  * fashion.
1458  * Note:
1459  *      - It is assumed that the socket was locked by the caller.
1460  *      - The routine does not block.
1461  *      - At present, there is no support for reading OOB data
1462  *        or for 'peeking' the socket using this routine
1463  *        (although both would be easy to implement).
1464  */
1465 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1466                   sk_read_actor_t recv_actor)
1467 {
1468         struct sk_buff *skb;
1469         struct tcp_sock *tp = tcp_sk(sk);
1470         u32 seq = tp->copied_seq;
1471         u32 offset;
1472         int copied = 0;
1473
1474         if (sk->sk_state == TCP_LISTEN)
1475                 return -ENOTCONN;
1476         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1477                 if (offset < skb->len) {
1478                         int used;
1479                         size_t len;
1480
1481                         len = skb->len - offset;
1482                         /* Stop reading if we hit a patch of urgent data */
1483                         if (tp->urg_data) {
1484                                 u32 urg_offset = tp->urg_seq - seq;
1485                                 if (urg_offset < len)
1486                                         len = urg_offset;
1487                                 if (!len)
1488                                         break;
1489                         }
1490                         used = recv_actor(desc, skb, offset, len);
1491                         if (used <= 0) {
1492                                 if (!copied)
1493                                         copied = used;
1494                                 break;
1495                         } else if (used <= len) {
1496                                 seq += used;
1497                                 copied += used;
1498                                 offset += used;
1499                         }
1500                         /* If recv_actor drops the lock (e.g. TCP splice
1501                          * receive) the skb pointer might be invalid when
1502                          * getting here: tcp_collapse might have deleted it
1503                          * while aggregating skbs from the socket queue.
1504                          */
1505                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1506                         if (!skb)
1507                                 break;
1508                         /* TCP coalescing might have appended data to the skb.
1509                          * Try to splice more frags
1510                          */
1511                         if (offset + 1 != skb->len)
1512                                 continue;
1513                 }
1514                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1515                         sk_eat_skb(sk, skb);
1516                         ++seq;
1517                         break;
1518                 }
1519                 sk_eat_skb(sk, skb);
1520                 if (!desc->count)
1521                         break;
1522                 tp->copied_seq = seq;
1523         }
1524         tp->copied_seq = seq;
1525
1526         tcp_rcv_space_adjust(sk);
1527
1528         /* Clean up data we have read: This will do ACK frames. */
1529         if (copied > 0) {
1530                 tcp_recv_skb(sk, seq, &offset);
1531                 tcp_cleanup_rbuf(sk, copied);
1532         }
1533         return copied;
1534 }
1535 EXPORT_SYMBOL(tcp_read_sock);
1536
1537 /*
1538  *      This routine copies from a sock struct into the user buffer.
1539  *
1540  *      Technical note: in 2.3 we work on _locked_ socket, so that
1541  *      tricks with *seq access order and skb->users are not required.
1542  *      Probably, code can be easily improved even more.
1543  */
1544
1545 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1546                 int flags, int *addr_len)
1547 {
1548         struct tcp_sock *tp = tcp_sk(sk);
1549         int copied = 0;
1550         u32 peek_seq;
1551         u32 *seq;
1552         unsigned long used;
1553         int err;
1554         int target;             /* Read at least this many bytes */
1555         long timeo;
1556         struct task_struct *user_recv = NULL;
1557         struct sk_buff *skb;
1558         u32 urg_hole = 0;
1559
1560         if (unlikely(flags & MSG_ERRQUEUE))
1561                 return inet_recv_error(sk, msg, len, addr_len);
1562
1563         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1564             (sk->sk_state == TCP_ESTABLISHED))
1565                 sk_busy_loop(sk, nonblock);
1566
1567         lock_sock(sk);
1568
1569         err = -ENOTCONN;
1570         if (sk->sk_state == TCP_LISTEN)
1571                 goto out;
1572
1573         timeo = sock_rcvtimeo(sk, nonblock);
1574
1575         /* Urgent data needs to be handled specially. */
1576         if (flags & MSG_OOB)
1577                 goto recv_urg;
1578
1579         if (unlikely(tp->repair)) {
1580                 err = -EPERM;
1581                 if (!(flags & MSG_PEEK))
1582                         goto out;
1583
1584                 if (tp->repair_queue == TCP_SEND_QUEUE)
1585                         goto recv_sndq;
1586
1587                 err = -EINVAL;
1588                 if (tp->repair_queue == TCP_NO_QUEUE)
1589                         goto out;
1590
1591                 /* 'common' recv queue MSG_PEEK-ing */
1592         }
1593
1594         seq = &tp->copied_seq;
1595         if (flags & MSG_PEEK) {
1596                 peek_seq = tp->copied_seq;
1597                 seq = &peek_seq;
1598         }
1599
1600         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1601
1602         do {
1603                 u32 offset;
1604
1605                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1606                 if (tp->urg_data && tp->urg_seq == *seq) {
1607                         if (copied)
1608                                 break;
1609                         if (signal_pending(current)) {
1610                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1611                                 break;
1612                         }
1613                 }
1614
1615                 /* Next get a buffer. */
1616
1617                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1618                         /* Now that we have two receive queues this
1619                          * shouldn't happen.
1620                          */
1621                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1622                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1623                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1624                                  flags))
1625                                 break;
1626
1627                         offset = *seq - TCP_SKB_CB(skb)->seq;
1628                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1629                                 offset--;
1630                         if (offset < skb->len)
1631                                 goto found_ok_skb;
1632                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1633                                 goto found_fin_ok;
1634                         WARN(!(flags & MSG_PEEK),
1635                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1636                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1637                 }
1638
1639                 /* Well, if we have backlog, try to process it now yet. */
1640
1641                 if (copied >= target && !sk->sk_backlog.tail)
1642                         break;
1643
1644                 if (copied) {
1645                         if (sk->sk_err ||
1646                             sk->sk_state == TCP_CLOSE ||
1647                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1648                             !timeo ||
1649                             signal_pending(current))
1650                                 break;
1651                 } else {
1652                         if (sock_flag(sk, SOCK_DONE))
1653                                 break;
1654
1655                         if (sk->sk_err) {
1656                                 copied = sock_error(sk);
1657                                 break;
1658                         }
1659
1660                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1661                                 break;
1662
1663                         if (sk->sk_state == TCP_CLOSE) {
1664                                 if (!sock_flag(sk, SOCK_DONE)) {
1665                                         /* This occurs when user tries to read
1666                                          * from never connected socket.
1667                                          */
1668                                         copied = -ENOTCONN;
1669                                         break;
1670                                 }
1671                                 break;
1672                         }
1673
1674                         if (!timeo) {
1675                                 copied = -EAGAIN;
1676                                 break;
1677                         }
1678
1679                         if (signal_pending(current)) {
1680                                 copied = sock_intr_errno(timeo);
1681                                 break;
1682                         }
1683                 }
1684
1685                 tcp_cleanup_rbuf(sk, copied);
1686
1687                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1688                         /* Install new reader */
1689                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1690                                 user_recv = current;
1691                                 tp->ucopy.task = user_recv;
1692                                 tp->ucopy.msg = msg;
1693                         }
1694
1695                         tp->ucopy.len = len;
1696
1697                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1698                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1699
1700                         /* Ugly... If prequeue is not empty, we have to
1701                          * process it before releasing socket, otherwise
1702                          * order will be broken at second iteration.
1703                          * More elegant solution is required!!!
1704                          *
1705                          * Look: we have the following (pseudo)queues:
1706                          *
1707                          * 1. packets in flight
1708                          * 2. backlog
1709                          * 3. prequeue
1710                          * 4. receive_queue
1711                          *
1712                          * Each queue can be processed only if the next ones
1713                          * are empty. At this point we have empty receive_queue.
1714                          * But prequeue _can_ be not empty after 2nd iteration,
1715                          * when we jumped to start of loop because backlog
1716                          * processing added something to receive_queue.
1717                          * We cannot release_sock(), because backlog contains
1718                          * packets arrived _after_ prequeued ones.
1719                          *
1720                          * Shortly, algorithm is clear --- to process all
1721                          * the queues in order. We could make it more directly,
1722                          * requeueing packets from backlog to prequeue, if
1723                          * is not empty. It is more elegant, but eats cycles,
1724                          * unfortunately.
1725                          */
1726                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1727                                 goto do_prequeue;
1728
1729                         /* __ Set realtime policy in scheduler __ */
1730                 }
1731
1732                 if (copied >= target) {
1733                         /* Do not sleep, just process backlog. */
1734                         release_sock(sk);
1735                         lock_sock(sk);
1736                 } else
1737                         sk_wait_data(sk, &timeo);
1738
1739                 if (user_recv) {
1740                         int chunk;
1741
1742                         /* __ Restore normal policy in scheduler __ */
1743
1744                         if ((chunk = len - tp->ucopy.len) != 0) {
1745                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1746                                 len -= chunk;
1747                                 copied += chunk;
1748                         }
1749
1750                         if (tp->rcv_nxt == tp->copied_seq &&
1751                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1752 do_prequeue:
1753                                 tcp_prequeue_process(sk);
1754
1755                                 if ((chunk = len - tp->ucopy.len) != 0) {
1756                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1757                                         len -= chunk;
1758                                         copied += chunk;
1759                                 }
1760                         }
1761                 }
1762                 if ((flags & MSG_PEEK) &&
1763                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1764                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1765                                             current->comm,
1766                                             task_pid_nr(current));
1767                         peek_seq = tp->copied_seq;
1768                 }
1769                 continue;
1770
1771         found_ok_skb:
1772                 /* Ok so how much can we use? */
1773                 used = skb->len - offset;
1774                 if (len < used)
1775                         used = len;
1776
1777                 /* Do we have urgent data here? */
1778                 if (tp->urg_data) {
1779                         u32 urg_offset = tp->urg_seq - *seq;
1780                         if (urg_offset < used) {
1781                                 if (!urg_offset) {
1782                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1783                                                 ++*seq;
1784                                                 urg_hole++;
1785                                                 offset++;
1786                                                 used--;
1787                                                 if (!used)
1788                                                         goto skip_copy;
1789                                         }
1790                                 } else
1791                                         used = urg_offset;
1792                         }
1793                 }
1794
1795                 if (!(flags & MSG_TRUNC)) {
1796                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1797                         if (err) {
1798                                 /* Exception. Bailout! */
1799                                 if (!copied)
1800                                         copied = -EFAULT;
1801                                 break;
1802                         }
1803                 }
1804
1805                 *seq += used;
1806                 copied += used;
1807                 len -= used;
1808
1809                 tcp_rcv_space_adjust(sk);
1810
1811 skip_copy:
1812                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1813                         tp->urg_data = 0;
1814                         tcp_fast_path_check(sk);
1815                 }
1816                 if (used + offset < skb->len)
1817                         continue;
1818
1819                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1820                         goto found_fin_ok;
1821                 if (!(flags & MSG_PEEK))
1822                         sk_eat_skb(sk, skb);
1823                 continue;
1824
1825         found_fin_ok:
1826                 /* Process the FIN. */
1827                 ++*seq;
1828                 if (!(flags & MSG_PEEK))
1829                         sk_eat_skb(sk, skb);
1830                 break;
1831         } while (len > 0);
1832
1833         if (user_recv) {
1834                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1835                         int chunk;
1836
1837                         tp->ucopy.len = copied > 0 ? len : 0;
1838
1839                         tcp_prequeue_process(sk);
1840
1841                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1842                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1843                                 len -= chunk;
1844                                 copied += chunk;
1845                         }
1846                 }
1847
1848                 tp->ucopy.task = NULL;
1849                 tp->ucopy.len = 0;
1850         }
1851
1852         /* According to UNIX98, msg_name/msg_namelen are ignored
1853          * on connected socket. I was just happy when found this 8) --ANK
1854          */
1855
1856         /* Clean up data we have read: This will do ACK frames. */
1857         tcp_cleanup_rbuf(sk, copied);
1858
1859         release_sock(sk);
1860         return copied;
1861
1862 out:
1863         release_sock(sk);
1864         return err;
1865
1866 recv_urg:
1867         err = tcp_recv_urg(sk, msg, len, flags);
1868         goto out;
1869
1870 recv_sndq:
1871         err = tcp_peek_sndq(sk, msg, len);
1872         goto out;
1873 }
1874 EXPORT_SYMBOL(tcp_recvmsg);
1875
1876 void tcp_set_state(struct sock *sk, int state)
1877 {
1878         int oldstate = sk->sk_state;
1879
1880         switch (state) {
1881         case TCP_ESTABLISHED:
1882                 if (oldstate != TCP_ESTABLISHED)
1883                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1884                 break;
1885
1886         case TCP_CLOSE:
1887                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1888                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1889
1890                 sk->sk_prot->unhash(sk);
1891                 if (inet_csk(sk)->icsk_bind_hash &&
1892                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1893                         inet_put_port(sk);
1894                 /* fall through */
1895         default:
1896                 if (oldstate == TCP_ESTABLISHED)
1897                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1898         }
1899
1900         /* Change state AFTER socket is unhashed to avoid closed
1901          * socket sitting in hash tables.
1902          */
1903         sk->sk_state = state;
1904
1905 #ifdef STATE_TRACE
1906         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1907 #endif
1908 }
1909 EXPORT_SYMBOL_GPL(tcp_set_state);
1910
1911 /*
1912  *      State processing on a close. This implements the state shift for
1913  *      sending our FIN frame. Note that we only send a FIN for some
1914  *      states. A shutdown() may have already sent the FIN, or we may be
1915  *      closed.
1916  */
1917
1918 static const unsigned char new_state[16] = {
1919   /* current state:        new state:      action:      */
1920   [0 /* (Invalid) */]   = TCP_CLOSE,
1921   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1922   [TCP_SYN_SENT]        = TCP_CLOSE,
1923   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1924   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
1925   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
1926   [TCP_TIME_WAIT]       = TCP_CLOSE,
1927   [TCP_CLOSE]           = TCP_CLOSE,
1928   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
1929   [TCP_LAST_ACK]        = TCP_LAST_ACK,
1930   [TCP_LISTEN]          = TCP_CLOSE,
1931   [TCP_CLOSING]         = TCP_CLOSING,
1932   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
1933 };
1934
1935 static int tcp_close_state(struct sock *sk)
1936 {
1937         int next = (int)new_state[sk->sk_state];
1938         int ns = next & TCP_STATE_MASK;
1939
1940         tcp_set_state(sk, ns);
1941
1942         return next & TCP_ACTION_FIN;
1943 }
1944
1945 /*
1946  *      Shutdown the sending side of a connection. Much like close except
1947  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1948  */
1949
1950 void tcp_shutdown(struct sock *sk, int how)
1951 {
1952         /*      We need to grab some memory, and put together a FIN,
1953          *      and then put it into the queue to be sent.
1954          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1955          */
1956         if (!(how & SEND_SHUTDOWN))
1957                 return;
1958
1959         /* If we've already sent a FIN, or it's a closed state, skip this. */
1960         if ((1 << sk->sk_state) &
1961             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1962              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1963                 /* Clear out any half completed packets.  FIN if needed. */
1964                 if (tcp_close_state(sk))
1965                         tcp_send_fin(sk);
1966         }
1967 }
1968 EXPORT_SYMBOL(tcp_shutdown);
1969
1970 bool tcp_check_oom(struct sock *sk, int shift)
1971 {
1972         bool too_many_orphans, out_of_socket_memory;
1973
1974         too_many_orphans = tcp_too_many_orphans(sk, shift);
1975         out_of_socket_memory = tcp_out_of_memory(sk);
1976
1977         if (too_many_orphans)
1978                 net_info_ratelimited("too many orphaned sockets\n");
1979         if (out_of_socket_memory)
1980                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
1981         return too_many_orphans || out_of_socket_memory;
1982 }
1983
1984 void tcp_close(struct sock *sk, long timeout)
1985 {
1986         struct sk_buff *skb;
1987         int data_was_unread = 0;
1988         int state;
1989
1990         lock_sock(sk);
1991         sk->sk_shutdown = SHUTDOWN_MASK;
1992
1993         if (sk->sk_state == TCP_LISTEN) {
1994                 tcp_set_state(sk, TCP_CLOSE);
1995
1996                 /* Special case. */
1997                 inet_csk_listen_stop(sk);
1998
1999                 goto adjudge_to_death;
2000         }
2001
2002         /*  We need to flush the recv. buffs.  We do this only on the
2003          *  descriptor close, not protocol-sourced closes, because the
2004          *  reader process may not have drained the data yet!
2005          */
2006         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2007                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2008
2009                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2010                         len--;
2011                 data_was_unread += len;
2012                 __kfree_skb(skb);
2013         }
2014
2015         sk_mem_reclaim(sk);
2016
2017         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2018         if (sk->sk_state == TCP_CLOSE)
2019                 goto adjudge_to_death;
2020
2021         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2022          * data was lost. To witness the awful effects of the old behavior of
2023          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2024          * GET in an FTP client, suspend the process, wait for the client to
2025          * advertise a zero window, then kill -9 the FTP client, wheee...
2026          * Note: timeout is always zero in such a case.
2027          */
2028         if (unlikely(tcp_sk(sk)->repair)) {
2029                 sk->sk_prot->disconnect(sk, 0);
2030         } else if (data_was_unread) {
2031                 /* Unread data was tossed, zap the connection. */
2032                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2033                 tcp_set_state(sk, TCP_CLOSE);
2034                 tcp_send_active_reset(sk, sk->sk_allocation);
2035         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2036                 /* Check zero linger _after_ checking for unread data. */
2037                 sk->sk_prot->disconnect(sk, 0);
2038                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2039         } else if (tcp_close_state(sk)) {
2040                 /* We FIN if the application ate all the data before
2041                  * zapping the connection.
2042                  */
2043
2044                 /* RED-PEN. Formally speaking, we have broken TCP state
2045                  * machine. State transitions:
2046                  *
2047                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2048                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2049                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2050                  *
2051                  * are legal only when FIN has been sent (i.e. in window),
2052                  * rather than queued out of window. Purists blame.
2053                  *
2054                  * F.e. "RFC state" is ESTABLISHED,
2055                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2056                  *
2057                  * The visible declinations are that sometimes
2058                  * we enter time-wait state, when it is not required really
2059                  * (harmless), do not send active resets, when they are
2060                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2061                  * they look as CLOSING or LAST_ACK for Linux)
2062                  * Probably, I missed some more holelets.
2063                  *                                              --ANK
2064                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2065                  * in a single packet! (May consider it later but will
2066                  * probably need API support or TCP_CORK SYN-ACK until
2067                  * data is written and socket is closed.)
2068                  */
2069                 tcp_send_fin(sk);
2070         }
2071
2072         sk_stream_wait_close(sk, timeout);
2073
2074 adjudge_to_death:
2075         state = sk->sk_state;
2076         sock_hold(sk);
2077         sock_orphan(sk);
2078
2079         /* It is the last release_sock in its life. It will remove backlog. */
2080         release_sock(sk);
2081
2082
2083         /* Now socket is owned by kernel and we acquire BH lock
2084            to finish close. No need to check for user refs.
2085          */
2086         local_bh_disable();
2087         bh_lock_sock(sk);
2088         WARN_ON(sock_owned_by_user(sk));
2089
2090         percpu_counter_inc(sk->sk_prot->orphan_count);
2091
2092         /* Have we already been destroyed by a softirq or backlog? */
2093         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2094                 goto out;
2095
2096         /*      This is a (useful) BSD violating of the RFC. There is a
2097          *      problem with TCP as specified in that the other end could
2098          *      keep a socket open forever with no application left this end.
2099          *      We use a 1 minute timeout (about the same as BSD) then kill
2100          *      our end. If they send after that then tough - BUT: long enough
2101          *      that we won't make the old 4*rto = almost no time - whoops
2102          *      reset mistake.
2103          *
2104          *      Nope, it was not mistake. It is really desired behaviour
2105          *      f.e. on http servers, when such sockets are useless, but
2106          *      consume significant resources. Let's do it with special
2107          *      linger2 option.                                 --ANK
2108          */
2109
2110         if (sk->sk_state == TCP_FIN_WAIT2) {
2111                 struct tcp_sock *tp = tcp_sk(sk);
2112                 if (tp->linger2 < 0) {
2113                         tcp_set_state(sk, TCP_CLOSE);
2114                         tcp_send_active_reset(sk, GFP_ATOMIC);
2115                         NET_INC_STATS_BH(sock_net(sk),
2116                                         LINUX_MIB_TCPABORTONLINGER);
2117                 } else {
2118                         const int tmo = tcp_fin_time(sk);
2119
2120                         if (tmo > TCP_TIMEWAIT_LEN) {
2121                                 inet_csk_reset_keepalive_timer(sk,
2122                                                 tmo - TCP_TIMEWAIT_LEN);
2123                         } else {
2124                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2125                                 goto out;
2126                         }
2127                 }
2128         }
2129         if (sk->sk_state != TCP_CLOSE) {
2130                 sk_mem_reclaim(sk);
2131                 if (tcp_check_oom(sk, 0)) {
2132                         tcp_set_state(sk, TCP_CLOSE);
2133                         tcp_send_active_reset(sk, GFP_ATOMIC);
2134                         NET_INC_STATS_BH(sock_net(sk),
2135                                         LINUX_MIB_TCPABORTONMEMORY);
2136                 }
2137         }
2138
2139         if (sk->sk_state == TCP_CLOSE) {
2140                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2141                 /* We could get here with a non-NULL req if the socket is
2142                  * aborted (e.g., closed with unread data) before 3WHS
2143                  * finishes.
2144                  */
2145                 if (req)
2146                         reqsk_fastopen_remove(sk, req, false);
2147                 inet_csk_destroy_sock(sk);
2148         }
2149         /* Otherwise, socket is reprieved until protocol close. */
2150
2151 out:
2152         bh_unlock_sock(sk);
2153         local_bh_enable();
2154         sock_put(sk);
2155 }
2156 EXPORT_SYMBOL(tcp_close);
2157
2158 /* These states need RST on ABORT according to RFC793 */
2159
2160 static inline bool tcp_need_reset(int state)
2161 {
2162         return (1 << state) &
2163                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2164                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2165 }
2166
2167 int tcp_disconnect(struct sock *sk, int flags)
2168 {
2169         struct inet_sock *inet = inet_sk(sk);
2170         struct inet_connection_sock *icsk = inet_csk(sk);
2171         struct tcp_sock *tp = tcp_sk(sk);
2172         int err = 0;
2173         int old_state = sk->sk_state;
2174
2175         if (old_state != TCP_CLOSE)
2176                 tcp_set_state(sk, TCP_CLOSE);
2177
2178         /* ABORT function of RFC793 */
2179         if (old_state == TCP_LISTEN) {
2180                 inet_csk_listen_stop(sk);
2181         } else if (unlikely(tp->repair)) {
2182                 sk->sk_err = ECONNABORTED;
2183         } else if (tcp_need_reset(old_state) ||
2184                    (tp->snd_nxt != tp->write_seq &&
2185                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2186                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2187                  * states
2188                  */
2189                 tcp_send_active_reset(sk, gfp_any());
2190                 sk->sk_err = ECONNRESET;
2191         } else if (old_state == TCP_SYN_SENT)
2192                 sk->sk_err = ECONNRESET;
2193
2194         tcp_clear_xmit_timers(sk);
2195         __skb_queue_purge(&sk->sk_receive_queue);
2196         tcp_write_queue_purge(sk);
2197         __skb_queue_purge(&tp->out_of_order_queue);
2198
2199         inet->inet_dport = 0;
2200
2201         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2202                 inet_reset_saddr(sk);
2203
2204         sk->sk_shutdown = 0;
2205         sock_reset_flag(sk, SOCK_DONE);
2206         tp->srtt_us = 0;
2207         if ((tp->write_seq += tp->max_window + 2) == 0)
2208                 tp->write_seq = 1;
2209         icsk->icsk_backoff = 0;
2210         tp->snd_cwnd = 2;
2211         icsk->icsk_probes_out = 0;
2212         tp->packets_out = 0;
2213         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2214         tp->snd_cwnd_cnt = 0;
2215         tp->window_clamp = 0;
2216         tcp_set_ca_state(sk, TCP_CA_Open);
2217         tcp_clear_retrans(tp);
2218         inet_csk_delack_init(sk);
2219         tcp_init_send_head(sk);
2220         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2221         __sk_dst_reset(sk);
2222
2223         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2224
2225         sk->sk_error_report(sk);
2226         return err;
2227 }
2228 EXPORT_SYMBOL(tcp_disconnect);
2229
2230 void tcp_sock_destruct(struct sock *sk)
2231 {
2232         inet_sock_destruct(sk);
2233
2234         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2235 }
2236
2237 static inline bool tcp_can_repair_sock(const struct sock *sk)
2238 {
2239         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2240                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2241 }
2242
2243 static int tcp_repair_options_est(struct tcp_sock *tp,
2244                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2245 {
2246         struct tcp_repair_opt opt;
2247
2248         while (len >= sizeof(opt)) {
2249                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2250                         return -EFAULT;
2251
2252                 optbuf++;
2253                 len -= sizeof(opt);
2254
2255                 switch (opt.opt_code) {
2256                 case TCPOPT_MSS:
2257                         tp->rx_opt.mss_clamp = opt.opt_val;
2258                         break;
2259                 case TCPOPT_WINDOW:
2260                         {
2261                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2262                                 u16 rcv_wscale = opt.opt_val >> 16;
2263
2264                                 if (snd_wscale > 14 || rcv_wscale > 14)
2265                                         return -EFBIG;
2266
2267                                 tp->rx_opt.snd_wscale = snd_wscale;
2268                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2269                                 tp->rx_opt.wscale_ok = 1;
2270                         }
2271                         break;
2272                 case TCPOPT_SACK_PERM:
2273                         if (opt.opt_val != 0)
2274                                 return -EINVAL;
2275
2276                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2277                         if (sysctl_tcp_fack)
2278                                 tcp_enable_fack(tp);
2279                         break;
2280                 case TCPOPT_TIMESTAMP:
2281                         if (opt.opt_val != 0)
2282                                 return -EINVAL;
2283
2284                         tp->rx_opt.tstamp_ok = 1;
2285                         break;
2286                 }
2287         }
2288
2289         return 0;
2290 }
2291
2292 /*
2293  *      Socket option code for TCP.
2294  */
2295 static int do_tcp_setsockopt(struct sock *sk, int level,
2296                 int optname, char __user *optval, unsigned int optlen)
2297 {
2298         struct tcp_sock *tp = tcp_sk(sk);
2299         struct inet_connection_sock *icsk = inet_csk(sk);
2300         int val;
2301         int err = 0;
2302
2303         /* These are data/string values, all the others are ints */
2304         switch (optname) {
2305         case TCP_CONGESTION: {
2306                 char name[TCP_CA_NAME_MAX];
2307
2308                 if (optlen < 1)
2309                         return -EINVAL;
2310
2311                 val = strncpy_from_user(name, optval,
2312                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2313                 if (val < 0)
2314                         return -EFAULT;
2315                 name[val] = 0;
2316
2317                 lock_sock(sk);
2318                 err = tcp_set_congestion_control(sk, name);
2319                 release_sock(sk);
2320                 return err;
2321         }
2322         default:
2323                 /* fallthru */
2324                 break;
2325         }
2326
2327         if (optlen < sizeof(int))
2328                 return -EINVAL;
2329
2330         if (get_user(val, (int __user *)optval))
2331                 return -EFAULT;
2332
2333         lock_sock(sk);
2334
2335         switch (optname) {
2336         case TCP_MAXSEG:
2337                 /* Values greater than interface MTU won't take effect. However
2338                  * at the point when this call is done we typically don't yet
2339                  * know which interface is going to be used */
2340                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2341                         err = -EINVAL;
2342                         break;
2343                 }
2344                 tp->rx_opt.user_mss = val;
2345                 break;
2346
2347         case TCP_NODELAY:
2348                 if (val) {
2349                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2350                          * this option on corked socket is remembered, but
2351                          * it is not activated until cork is cleared.
2352                          *
2353                          * However, when TCP_NODELAY is set we make
2354                          * an explicit push, which overrides even TCP_CORK
2355                          * for currently queued segments.
2356                          */
2357                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2358                         tcp_push_pending_frames(sk);
2359                 } else {
2360                         tp->nonagle &= ~TCP_NAGLE_OFF;
2361                 }
2362                 break;
2363
2364         case TCP_THIN_LINEAR_TIMEOUTS:
2365                 if (val < 0 || val > 1)
2366                         err = -EINVAL;
2367                 else
2368                         tp->thin_lto = val;
2369                 break;
2370
2371         case TCP_THIN_DUPACK:
2372                 if (val < 0 || val > 1)
2373                         err = -EINVAL;
2374                 else {
2375                         tp->thin_dupack = val;
2376                         if (tp->thin_dupack)
2377                                 tcp_disable_early_retrans(tp);
2378                 }
2379                 break;
2380
2381         case TCP_REPAIR:
2382                 if (!tcp_can_repair_sock(sk))
2383                         err = -EPERM;
2384                 else if (val == 1) {
2385                         tp->repair = 1;
2386                         sk->sk_reuse = SK_FORCE_REUSE;
2387                         tp->repair_queue = TCP_NO_QUEUE;
2388                 } else if (val == 0) {
2389                         tp->repair = 0;
2390                         sk->sk_reuse = SK_NO_REUSE;
2391                         tcp_send_window_probe(sk);
2392                 } else
2393                         err = -EINVAL;
2394
2395                 break;
2396
2397         case TCP_REPAIR_QUEUE:
2398                 if (!tp->repair)
2399                         err = -EPERM;
2400                 else if (val < TCP_QUEUES_NR)
2401                         tp->repair_queue = val;
2402                 else
2403                         err = -EINVAL;
2404                 break;
2405
2406         case TCP_QUEUE_SEQ:
2407                 if (sk->sk_state != TCP_CLOSE)
2408                         err = -EPERM;
2409                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2410                         tp->write_seq = val;
2411                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2412                         tp->rcv_nxt = val;
2413                 else
2414                         err = -EINVAL;
2415                 break;
2416
2417         case TCP_REPAIR_OPTIONS:
2418                 if (!tp->repair)
2419                         err = -EINVAL;
2420                 else if (sk->sk_state == TCP_ESTABLISHED)
2421                         err = tcp_repair_options_est(tp,
2422                                         (struct tcp_repair_opt __user *)optval,
2423                                         optlen);
2424                 else
2425                         err = -EPERM;
2426                 break;
2427
2428         case TCP_CORK:
2429                 /* When set indicates to always queue non-full frames.
2430                  * Later the user clears this option and we transmit
2431                  * any pending partial frames in the queue.  This is
2432                  * meant to be used alongside sendfile() to get properly
2433                  * filled frames when the user (for example) must write
2434                  * out headers with a write() call first and then use
2435                  * sendfile to send out the data parts.
2436                  *
2437                  * TCP_CORK can be set together with TCP_NODELAY and it is
2438                  * stronger than TCP_NODELAY.
2439                  */
2440                 if (val) {
2441                         tp->nonagle |= TCP_NAGLE_CORK;
2442                 } else {
2443                         tp->nonagle &= ~TCP_NAGLE_CORK;
2444                         if (tp->nonagle&TCP_NAGLE_OFF)
2445                                 tp->nonagle |= TCP_NAGLE_PUSH;
2446                         tcp_push_pending_frames(sk);
2447                 }
2448                 break;
2449
2450         case TCP_KEEPIDLE:
2451                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2452                         err = -EINVAL;
2453                 else {
2454                         tp->keepalive_time = val * HZ;
2455                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2456                             !((1 << sk->sk_state) &
2457                               (TCPF_CLOSE | TCPF_LISTEN))) {
2458                                 u32 elapsed = keepalive_time_elapsed(tp);
2459                                 if (tp->keepalive_time > elapsed)
2460                                         elapsed = tp->keepalive_time - elapsed;
2461                                 else
2462                                         elapsed = 0;
2463                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2464                         }
2465                 }
2466                 break;
2467         case TCP_KEEPINTVL:
2468                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2469                         err = -EINVAL;
2470                 else
2471                         tp->keepalive_intvl = val * HZ;
2472                 break;
2473         case TCP_KEEPCNT:
2474                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2475                         err = -EINVAL;
2476                 else
2477                         tp->keepalive_probes = val;
2478                 break;
2479         case TCP_SYNCNT:
2480                 if (val < 1 || val > MAX_TCP_SYNCNT)
2481                         err = -EINVAL;
2482                 else
2483                         icsk->icsk_syn_retries = val;
2484                 break;
2485
2486         case TCP_LINGER2:
2487                 if (val < 0)
2488                         tp->linger2 = -1;
2489                 else if (val > sysctl_tcp_fin_timeout / HZ)
2490                         tp->linger2 = 0;
2491                 else
2492                         tp->linger2 = val * HZ;
2493                 break;
2494
2495         case TCP_DEFER_ACCEPT:
2496                 /* Translate value in seconds to number of retransmits */
2497                 icsk->icsk_accept_queue.rskq_defer_accept =
2498                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2499                                         TCP_RTO_MAX / HZ);
2500                 break;
2501
2502         case TCP_WINDOW_CLAMP:
2503                 if (!val) {
2504                         if (sk->sk_state != TCP_CLOSE) {
2505                                 err = -EINVAL;
2506                                 break;
2507                         }
2508                         tp->window_clamp = 0;
2509                 } else
2510                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2511                                                 SOCK_MIN_RCVBUF / 2 : val;
2512                 break;
2513
2514         case TCP_QUICKACK:
2515                 if (!val) {
2516                         icsk->icsk_ack.pingpong = 1;
2517                 } else {
2518                         icsk->icsk_ack.pingpong = 0;
2519                         if ((1 << sk->sk_state) &
2520                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2521                             inet_csk_ack_scheduled(sk)) {
2522                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2523                                 tcp_cleanup_rbuf(sk, 1);
2524                                 if (!(val & 1))
2525                                         icsk->icsk_ack.pingpong = 1;
2526                         }
2527                 }
2528                 break;
2529
2530 #ifdef CONFIG_TCP_MD5SIG
2531         case TCP_MD5SIG:
2532                 /* Read the IP->Key mappings from userspace */
2533                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2534                 break;
2535 #endif
2536         case TCP_USER_TIMEOUT:
2537                 /* Cap the max time in ms TCP will retry or probe the window
2538                  * before giving up and aborting (ETIMEDOUT) a connection.
2539                  */
2540                 if (val < 0)
2541                         err = -EINVAL;
2542                 else
2543                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2544                 break;
2545
2546         case TCP_FASTOPEN:
2547                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2548                     TCPF_LISTEN))) {
2549                         tcp_fastopen_init_key_once(true);
2550
2551                         err = fastopen_init_queue(sk, val);
2552                 } else {
2553                         err = -EINVAL;
2554                 }
2555                 break;
2556         case TCP_TIMESTAMP:
2557                 if (!tp->repair)
2558                         err = -EPERM;
2559                 else
2560                         tp->tsoffset = val - tcp_time_stamp;
2561                 break;
2562         case TCP_NOTSENT_LOWAT:
2563                 tp->notsent_lowat = val;
2564                 sk->sk_write_space(sk);
2565                 break;
2566         default:
2567                 err = -ENOPROTOOPT;
2568                 break;
2569         }
2570
2571         release_sock(sk);
2572         return err;
2573 }
2574
2575 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2576                    unsigned int optlen)
2577 {
2578         const struct inet_connection_sock *icsk = inet_csk(sk);
2579
2580         if (level != SOL_TCP)
2581                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2582                                                      optval, optlen);
2583         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2584 }
2585 EXPORT_SYMBOL(tcp_setsockopt);
2586
2587 #ifdef CONFIG_COMPAT
2588 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2589                           char __user *optval, unsigned int optlen)
2590 {
2591         if (level != SOL_TCP)
2592                 return inet_csk_compat_setsockopt(sk, level, optname,
2593                                                   optval, optlen);
2594         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2595 }
2596 EXPORT_SYMBOL(compat_tcp_setsockopt);
2597 #endif
2598
2599 /* Return information about state of tcp endpoint in API format. */
2600 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2601 {
2602         const struct tcp_sock *tp = tcp_sk(sk);
2603         const struct inet_connection_sock *icsk = inet_csk(sk);
2604         u32 now = tcp_time_stamp;
2605         unsigned int start;
2606         u32 rate;
2607
2608         memset(info, 0, sizeof(*info));
2609
2610         info->tcpi_state = sk->sk_state;
2611         info->tcpi_ca_state = icsk->icsk_ca_state;
2612         info->tcpi_retransmits = icsk->icsk_retransmits;
2613         info->tcpi_probes = icsk->icsk_probes_out;
2614         info->tcpi_backoff = icsk->icsk_backoff;
2615
2616         if (tp->rx_opt.tstamp_ok)
2617                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2618         if (tcp_is_sack(tp))
2619                 info->tcpi_options |= TCPI_OPT_SACK;
2620         if (tp->rx_opt.wscale_ok) {
2621                 info->tcpi_options |= TCPI_OPT_WSCALE;
2622                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2623                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2624         }
2625
2626         if (tp->ecn_flags & TCP_ECN_OK)
2627                 info->tcpi_options |= TCPI_OPT_ECN;
2628         if (tp->ecn_flags & TCP_ECN_SEEN)
2629                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2630         if (tp->syn_data_acked)
2631                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2632
2633         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2634         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2635         info->tcpi_snd_mss = tp->mss_cache;
2636         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2637
2638         if (sk->sk_state == TCP_LISTEN) {
2639                 info->tcpi_unacked = sk->sk_ack_backlog;
2640                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2641         } else {
2642                 info->tcpi_unacked = tp->packets_out;
2643                 info->tcpi_sacked = tp->sacked_out;
2644         }
2645         info->tcpi_lost = tp->lost_out;
2646         info->tcpi_retrans = tp->retrans_out;
2647         info->tcpi_fackets = tp->fackets_out;
2648
2649         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2650         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2651         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2652
2653         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2654         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2655         info->tcpi_rtt = tp->srtt_us >> 3;
2656         info->tcpi_rttvar = tp->mdev_us >> 2;
2657         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2658         info->tcpi_snd_cwnd = tp->snd_cwnd;
2659         info->tcpi_advmss = tp->advmss;
2660         info->tcpi_reordering = tp->reordering;
2661
2662         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2663         info->tcpi_rcv_space = tp->rcvq_space.space;
2664
2665         info->tcpi_total_retrans = tp->total_retrans;
2666
2667         rate = READ_ONCE(sk->sk_pacing_rate);
2668         info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2669
2670         rate = READ_ONCE(sk->sk_max_pacing_rate);
2671         info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2672
2673         do {
2674                 start = u64_stats_fetch_begin_irq(&tp->syncp);
2675                 info->tcpi_bytes_acked = tp->bytes_acked;
2676                 info->tcpi_bytes_received = tp->bytes_received;
2677         } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2678 }
2679 EXPORT_SYMBOL_GPL(tcp_get_info);
2680
2681 static int do_tcp_getsockopt(struct sock *sk, int level,
2682                 int optname, char __user *optval, int __user *optlen)
2683 {
2684         struct inet_connection_sock *icsk = inet_csk(sk);
2685         struct tcp_sock *tp = tcp_sk(sk);
2686         int val, len;
2687
2688         if (get_user(len, optlen))
2689                 return -EFAULT;
2690
2691         len = min_t(unsigned int, len, sizeof(int));
2692
2693         if (len < 0)
2694                 return -EINVAL;
2695
2696         switch (optname) {
2697         case TCP_MAXSEG:
2698                 val = tp->mss_cache;
2699                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2700                         val = tp->rx_opt.user_mss;
2701                 if (tp->repair)
2702                         val = tp->rx_opt.mss_clamp;
2703                 break;
2704         case TCP_NODELAY:
2705                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2706                 break;
2707         case TCP_CORK:
2708                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2709                 break;
2710         case TCP_KEEPIDLE:
2711                 val = keepalive_time_when(tp) / HZ;
2712                 break;
2713         case TCP_KEEPINTVL:
2714                 val = keepalive_intvl_when(tp) / HZ;
2715                 break;
2716         case TCP_KEEPCNT:
2717                 val = keepalive_probes(tp);
2718                 break;
2719         case TCP_SYNCNT:
2720                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2721                 break;
2722         case TCP_LINGER2:
2723                 val = tp->linger2;
2724                 if (val >= 0)
2725                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2726                 break;
2727         case TCP_DEFER_ACCEPT:
2728                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2729                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2730                 break;
2731         case TCP_WINDOW_CLAMP:
2732                 val = tp->window_clamp;
2733                 break;
2734         case TCP_INFO: {
2735                 struct tcp_info info;
2736
2737                 if (get_user(len, optlen))
2738                         return -EFAULT;
2739
2740                 tcp_get_info(sk, &info);
2741
2742                 len = min_t(unsigned int, len, sizeof(info));
2743                 if (put_user(len, optlen))
2744                         return -EFAULT;
2745                 if (copy_to_user(optval, &info, len))
2746                         return -EFAULT;
2747                 return 0;
2748         }
2749         case TCP_CC_INFO: {
2750                 const struct tcp_congestion_ops *ca_ops;
2751                 union tcp_cc_info info;
2752                 size_t sz = 0;
2753                 int attr;
2754
2755                 if (get_user(len, optlen))
2756                         return -EFAULT;
2757
2758                 ca_ops = icsk->icsk_ca_ops;
2759                 if (ca_ops && ca_ops->get_info)
2760                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2761
2762                 len = min_t(unsigned int, len, sz);
2763                 if (put_user(len, optlen))
2764                         return -EFAULT;
2765                 if (copy_to_user(optval, &info, len))
2766                         return -EFAULT;
2767                 return 0;
2768         }
2769         case TCP_QUICKACK:
2770                 val = !icsk->icsk_ack.pingpong;
2771                 break;
2772
2773         case TCP_CONGESTION:
2774                 if (get_user(len, optlen))
2775                         return -EFAULT;
2776                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2777                 if (put_user(len, optlen))
2778                         return -EFAULT;
2779                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2780                         return -EFAULT;
2781                 return 0;
2782
2783         case TCP_THIN_LINEAR_TIMEOUTS:
2784                 val = tp->thin_lto;
2785                 break;
2786         case TCP_THIN_DUPACK:
2787                 val = tp->thin_dupack;
2788                 break;
2789
2790         case TCP_REPAIR:
2791                 val = tp->repair;
2792                 break;
2793
2794         case TCP_REPAIR_QUEUE:
2795                 if (tp->repair)
2796                         val = tp->repair_queue;
2797                 else
2798                         return -EINVAL;
2799                 break;
2800
2801         case TCP_QUEUE_SEQ:
2802                 if (tp->repair_queue == TCP_SEND_QUEUE)
2803                         val = tp->write_seq;
2804                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2805                         val = tp->rcv_nxt;
2806                 else
2807                         return -EINVAL;
2808                 break;
2809
2810         case TCP_USER_TIMEOUT:
2811                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2812                 break;
2813
2814         case TCP_FASTOPEN:
2815                 if (icsk->icsk_accept_queue.fastopenq)
2816                         val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2817                 else
2818                         val = 0;
2819                 break;
2820
2821         case TCP_TIMESTAMP:
2822                 val = tcp_time_stamp + tp->tsoffset;
2823                 break;
2824         case TCP_NOTSENT_LOWAT:
2825                 val = tp->notsent_lowat;
2826                 break;
2827         default:
2828                 return -ENOPROTOOPT;
2829         }
2830
2831         if (put_user(len, optlen))
2832                 return -EFAULT;
2833         if (copy_to_user(optval, &val, len))
2834                 return -EFAULT;
2835         return 0;
2836 }
2837
2838 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2839                    int __user *optlen)
2840 {
2841         struct inet_connection_sock *icsk = inet_csk(sk);
2842
2843         if (level != SOL_TCP)
2844                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2845                                                      optval, optlen);
2846         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2847 }
2848 EXPORT_SYMBOL(tcp_getsockopt);
2849
2850 #ifdef CONFIG_COMPAT
2851 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2852                           char __user *optval, int __user *optlen)
2853 {
2854         if (level != SOL_TCP)
2855                 return inet_csk_compat_getsockopt(sk, level, optname,
2856                                                   optval, optlen);
2857         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2858 }
2859 EXPORT_SYMBOL(compat_tcp_getsockopt);
2860 #endif
2861
2862 #ifdef CONFIG_TCP_MD5SIG
2863 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2864 static DEFINE_MUTEX(tcp_md5sig_mutex);
2865 static bool tcp_md5sig_pool_populated = false;
2866
2867 static void __tcp_alloc_md5sig_pool(void)
2868 {
2869         int cpu;
2870
2871         for_each_possible_cpu(cpu) {
2872                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2873                         struct crypto_hash *hash;
2874
2875                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2876                         if (IS_ERR_OR_NULL(hash))
2877                                 return;
2878                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2879                 }
2880         }
2881         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2882          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2883          */
2884         smp_wmb();
2885         tcp_md5sig_pool_populated = true;
2886 }
2887
2888 bool tcp_alloc_md5sig_pool(void)
2889 {
2890         if (unlikely(!tcp_md5sig_pool_populated)) {
2891                 mutex_lock(&tcp_md5sig_mutex);
2892
2893                 if (!tcp_md5sig_pool_populated)
2894                         __tcp_alloc_md5sig_pool();
2895
2896                 mutex_unlock(&tcp_md5sig_mutex);
2897         }
2898         return tcp_md5sig_pool_populated;
2899 }
2900 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2901
2902
2903 /**
2904  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2905  *
2906  *      We use percpu structure, so if we succeed, we exit with preemption
2907  *      and BH disabled, to make sure another thread or softirq handling
2908  *      wont try to get same context.
2909  */
2910 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2911 {
2912         local_bh_disable();
2913
2914         if (tcp_md5sig_pool_populated) {
2915                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2916                 smp_rmb();
2917                 return this_cpu_ptr(&tcp_md5sig_pool);
2918         }
2919         local_bh_enable();
2920         return NULL;
2921 }
2922 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2923
2924 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2925                         const struct tcphdr *th)
2926 {
2927         struct scatterlist sg;
2928         struct tcphdr hdr;
2929         int err;
2930
2931         /* We are not allowed to change tcphdr, make a local copy */
2932         memcpy(&hdr, th, sizeof(hdr));
2933         hdr.check = 0;
2934
2935         /* options aren't included in the hash */
2936         sg_init_one(&sg, &hdr, sizeof(hdr));
2937         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2938         return err;
2939 }
2940 EXPORT_SYMBOL(tcp_md5_hash_header);
2941
2942 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2943                           const struct sk_buff *skb, unsigned int header_len)
2944 {
2945         struct scatterlist sg;
2946         const struct tcphdr *tp = tcp_hdr(skb);
2947         struct hash_desc *desc = &hp->md5_desc;
2948         unsigned int i;
2949         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
2950                                            skb_headlen(skb) - header_len : 0;
2951         const struct skb_shared_info *shi = skb_shinfo(skb);
2952         struct sk_buff *frag_iter;
2953
2954         sg_init_table(&sg, 1);
2955
2956         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2957         if (crypto_hash_update(desc, &sg, head_data_len))
2958                 return 1;
2959
2960         for (i = 0; i < shi->nr_frags; ++i) {
2961                 const struct skb_frag_struct *f = &shi->frags[i];
2962                 unsigned int offset = f->page_offset;
2963                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2964
2965                 sg_set_page(&sg, page, skb_frag_size(f),
2966                             offset_in_page(offset));
2967                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2968                         return 1;
2969         }
2970
2971         skb_walk_frags(skb, frag_iter)
2972                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2973                         return 1;
2974
2975         return 0;
2976 }
2977 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2978
2979 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2980 {
2981         struct scatterlist sg;
2982
2983         sg_init_one(&sg, key->key, key->keylen);
2984         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2985 }
2986 EXPORT_SYMBOL(tcp_md5_hash_key);
2987
2988 #endif
2989
2990 void tcp_done(struct sock *sk)
2991 {
2992         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2993
2994         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2995                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2996
2997         tcp_set_state(sk, TCP_CLOSE);
2998         tcp_clear_xmit_timers(sk);
2999         if (req)
3000                 reqsk_fastopen_remove(sk, req, false);
3001
3002         sk->sk_shutdown = SHUTDOWN_MASK;
3003
3004         if (!sock_flag(sk, SOCK_DEAD))
3005                 sk->sk_state_change(sk);
3006         else
3007                 inet_csk_destroy_sock(sk);
3008 }
3009 EXPORT_SYMBOL_GPL(tcp_done);
3010
3011 extern struct tcp_congestion_ops tcp_reno;
3012
3013 static __initdata unsigned long thash_entries;
3014 static int __init set_thash_entries(char *str)
3015 {
3016         ssize_t ret;
3017
3018         if (!str)
3019                 return 0;
3020
3021         ret = kstrtoul(str, 0, &thash_entries);
3022         if (ret)
3023                 return 0;
3024
3025         return 1;
3026 }
3027 __setup("thash_entries=", set_thash_entries);
3028
3029 static void __init tcp_init_mem(void)
3030 {
3031         unsigned long limit = nr_free_buffer_pages() / 8;
3032         limit = max(limit, 128UL);
3033         sysctl_tcp_mem[0] = limit / 4 * 3;
3034         sysctl_tcp_mem[1] = limit;
3035         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3036 }
3037
3038 void __init tcp_init(void)
3039 {
3040         unsigned long limit;
3041         int max_rshare, max_wshare, cnt;
3042         unsigned int i;
3043
3044         sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3045
3046         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3047         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3048         tcp_hashinfo.bind_bucket_cachep =
3049                 kmem_cache_create("tcp_bind_bucket",
3050                                   sizeof(struct inet_bind_bucket), 0,
3051                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3052
3053         /* Size and allocate the main established and bind bucket
3054          * hash tables.
3055          *
3056          * The methodology is similar to that of the buffer cache.
3057          */
3058         tcp_hashinfo.ehash =
3059                 alloc_large_system_hash("TCP established",
3060                                         sizeof(struct inet_ehash_bucket),
3061                                         thash_entries,
3062                                         17, /* one slot per 128 KB of memory */
3063                                         0,
3064                                         NULL,
3065                                         &tcp_hashinfo.ehash_mask,
3066                                         0,
3067                                         thash_entries ? 0 : 512 * 1024);
3068         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3069                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3070
3071         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3072                 panic("TCP: failed to alloc ehash_locks");
3073         tcp_hashinfo.bhash =
3074                 alloc_large_system_hash("TCP bind",
3075                                         sizeof(struct inet_bind_hashbucket),
3076                                         tcp_hashinfo.ehash_mask + 1,
3077                                         17, /* one slot per 128 KB of memory */
3078                                         0,
3079                                         &tcp_hashinfo.bhash_size,
3080                                         NULL,
3081                                         0,
3082                                         64 * 1024);
3083         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3084         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3085                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3086                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3087         }
3088
3089
3090         cnt = tcp_hashinfo.ehash_mask + 1;
3091
3092         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3093         sysctl_tcp_max_orphans = cnt / 2;
3094         sysctl_max_syn_backlog = max(128, cnt / 256);
3095
3096         tcp_init_mem();
3097         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3098         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3099         max_wshare = min(4UL*1024*1024, limit);
3100         max_rshare = min(6UL*1024*1024, limit);
3101
3102         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3103         sysctl_tcp_wmem[1] = 16*1024;
3104         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3105
3106         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3107         sysctl_tcp_rmem[1] = 87380;
3108         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3109
3110         pr_info("Hash tables configured (established %u bind %u)\n",
3111                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3112
3113         tcp_metrics_init();
3114         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3115         tcp_tasklet_init();
3116 }